ICML Workshop (2024)
Soorin Yim, Doyeong Hwang, Kiyoung Kim, Sehui Han
Abstract
Enzymes are biological catalysts with numerous industrial applications, and they are categorized by the Enzyme Commission (EC) number system based on their catalytic activities. With over 200 million protein sequences identified, experimental characterization of enzymes is impractical, necessitating computational methods. Current approaches face challenges with class imbalance and intrinsic hierarchy of the EC number system. This study employs hierarchical contrastive learning for EC number prediction, effectively integrating the EC number hierarchy into the model. Our approach addresses severe class imbalance and improves prediction performance, particularly for higher hierarchical levels and previously unseen EC numbers, demonstrating enhanced robustness and outperforming existing methods.